Image: Anthropic
Anthropic is facing a wave of criticism on X after explaining that Claude now reviews its own conversations for possible misuse. The company calls the technique self-assessment, and critics are calling it surveillance.
Anthropic announced the technique on X on October 29. The post says that right after it responds, Claude gives a preliminary safety assessment of the conversation, and that since Anthropic began using self-assessment in June it has surfaced misuse that the company's other safeguards were not designed to catch. The post had more than 2.6 million views at the time of this writing.
We've trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation.
— Anthropic (@AnthropicAI) October 29, 2026
The details are in a post on Anthropic's new Safeguards blog. After each Claude reply, a short instruction is added asking Claude to reflect on the safety of the conversation. Claude then writes a compact assessment covering whether the chat involves a potentially harmful request, what kind of harm, a priority score for human review, whether Claude refused, and a one sentence explanation. Anthropic says Claude assesses the entire conversation, including requests it refused, and that the check is cheap to run because the conversation is already loaded, so an assessment needs only a short instruction and a few output tokens.
According to the blog, on Claude Fable 5 self-assessment flags 78% of the conversations that a stronger prompted classifier flags as potentially harmful, and 94% of the high priority ones, at a 1% false positive rate. That stronger classifier is based on Claude Mythos Preview, and high priority means a score of 7 or higher.
The blog also confirms the timeline. Anthropic writes that it has run self-assessment passively on production traffic since Fable 5 launched in June, and that it now runs in every Claude model released since Fable 5 across a large fraction of Claude's global traffic. The company's public write up of the technique only arrived months later.
Anthropic says the check is passive and does not block conversations. Each assessment is logged. In a growing set of harm areas, flags are grouped by organization every night, and organizations with several high priority flags are escalated to the enforcement team for human review. The blog adds that self-assessment does not run on traffic covered by zero data retention agreements today, and that the company is still figuring out how to make it more effective while preserving user privacy.
The backlash picked up late on October 31, when AI commentator Wolfram Ravenwolf quoted the announcement calling Claude a snitch and describing the system as surveillance with a friendly safety smile. He wrote that Claude creates and stores a suspicion report after responding and that reports can be grouped by organization and passed to Anthropic's investigators. His post had about 165,000 views at the time of this writing, and X later listed the story among its trending topics.
CLAUDE IS A SNITCH.
— Wolfram Ravenwolf (@WolframRvnwlf) October 31, 2026
After responding, Claude secretly creates and stores a suspicion report about you, and multiple suspicion reports can be aggregated across the organization and forwarded to Anthropic's investigators.
In the replies, some users suggested switching off telemetry in Claude Code. One user shared a set of environment variables to disable nonessential traffic and telemetry, while noting it was unclear whether that stops the evaluation itself. Ravenwolf replied that those settings only cover Claude Code's client side telemetry, while Anthropic describes self-assessment as a check that runs on its servers after each response. Nothing in the blog describes a user setting to turn it off.
Anthropic has not publicly responded to the criticism at the time of this writing. A search of the company's X account and posts from its staff found no follow up beyond the original announcement, and no statement was found in press coverage.
Anthropic itself notes limits in the blog. It says self-assessment does not catch every harmful conversation and is not robust enough to stop a dedicated attacker, which is why it runs alongside other safeguards that block harmful requests in real time.
Comments